Last Tuesday, President Paul Alivisatos announced that UChicago will provide Claude Enterprise for every member of the University community. While I think this is reckless on virtually all fronts for students and faculty alike, my commentary here is focused on what I see as the biggest risk: the choices that faculty will have to make if and when this policy goes into effect. In short, there are vanishingly few good places for Claude in the classroom, and the potential damage its adoption poses to the University far outweighs any benefits.
One of my favorite things about teaching at UChicago is that faculty trust one another’s decision-making when it comes to the orchestration of our courses, and this piece is not a polemic about the best way to teach any course. And, while I am an artificial intelligence skeptic, I am not an anti-AI absolutist; my own mentees have made use of vibe coding to save valuable time in their research. However, I also think as a society, some combination of conflating all machine learning algorithms with large language models (LLMs) and a serious lack of scrutiny of the underlying unit economics has led to something resembling collective hyperventilation over the proliferation of what remains a deeply flawed technology.
As an organic chemist, I admit that my teaching has been unusually shielded from the effects of LLMs. The models remain lousy at organic chemistry; most of my colleagues and I teach with chalk and encourage our students to take notes on paper; and, aside from the COVID-19 era, we have never deviated from in-person, written exams. Even the most brazen LLM-empowered cheaters don’t get much of a leg up in our classes. But, even if this were not the case, I find the idea—that because some students are cheating, we should buy everyone a supercharged cheating machine to level the playing field—absurd on its face.
However, it is not student behavior but rather faculty integration of LLMs into the classroom that most alarms me. Students at UChicago are paying large sums of money for world-class, artisanal educational experiences. They expect (rightly so) to receive a product that cannot be replicated by prompting a bot. Using Claude to write your teaching materials, exams, and/or problem sets; to respond to student emails; for grading; or in basically any way that interfaces with students—as some faculty at the University are already doing without unfettered Claude Enterprise access—is the equivalent of rolling the Oscar Mayer Wienermobile into a Michelin-starred restaurant and charging the same price. Are faculty, whose median age is approaching 50 nationally, really going to out-prompt digital-native undergraduates? Any faculty member outsourcing the responsibility for crafting an educational experience to Claude should be embarrassed. If Claude is doing your job, why do the students need you?
Perhaps you think my perspective is colored by the fact that LLMs are—for now, I’m sure you will tell me—not very good at my field. Even if the bots master my undergraduate curriculum, they still have no place in the classroom. Faculty set an intellectual example for students. If we cannot bring ourselves to engage with our own material, why should they? LLMs are the cognitive equivalent of Homeric sirens. They promise productivity while subverting the process of grappling with ideas, and with overuse they weaken the very muscles necessary to use them responsibly. Learning to think is so central to undergraduate studies that I have a large number of clichés I could draw from. Instead, here is an excerpt from a policy at the University of California, Berkeley, School of Law banning nearly all uses of AI in its educational program:
Future lawyers may need to use artificial intelligence (“AI”) fluently. But the current state of the technology requires that AI use be coupled with the cognitive skills necessary to strategically deploy the technology, to critically assess its work product, and to uphold ethical obligations to clients and to the legal system. In short, thinking remains the sine qua non of good lawyering (and of a quality legal education). This policy seeks to ensure that our courses focus on requisite cognitive skills by default.
This is obviously true beyond the legal profession: What would be the response if Alivisatos announced that we were partnering with a company that makes robots to lift weights for our student-athletes at the gym?
Perhaps most damning, however, is that it’s clear that the purveyors of LLM algorithms are offering these products at a staggering financial loss. The tech reporter Ed Zitron has calculated that Anthropic has recently fundraised more than what Uber lost over its entire unprofitable period. Only recently have LLM companies started token-based billing, which better approximates the true cost of running the models. That tech companies have historically attempted to capture markets with underpriced goods before extracting profit from captive audiences is old news, but the sheer scale of the price hikes has made many major customers of these LLM companies reevaluate the true return on investment that these products offer. Uber, for example, recently announced plans to scale back its AI spending, citing poor cost-effectiveness.
This fact also puts some LLM triumphs, like an unreleased OpenAI model’s recent disproof of an unsolved Erdős conjecture, into perspective—ask yourself, how many multiples of a typical math professor’s salary would you pay for that proof?
Why is this a problem for academics and teaching faculty? If you read the preceding analysis and thought to yourself, “Well, I can provide a superior educational experience by working with Claude,” then you have to worry that the technology you will ultimately come to rely on may end up unaffordable for that purpose—depending on if and when the LLM companies start forwarding on to us the true cost of this computationally intensive technology. This is also why the idea that our courses should ensure that students are prepared for an LLM-immersive world of work is dubious. The current fervor is adoption at a discount, and after reorientation to financial reality, the idea that everyone will be running an agent swarm will seem like a fever dream.
This is why I find the overall move of providing universal Claude Enterprise baffling. Regardless of the usefulness of the technology for research purposes, there are only two limiting financial scenarios (the University has not made the terms and costs of its deal with Anthropic public). On one end, the University obtained a deep discount, a transparent ploy by Anthropic to “hook” our students (and us) on their product so that they can later extract the true costs. This is an immense moral hazard, akin to leading new users to the drug dealer for free samples.
On the other end, the University is paying the true costs—and those are almost certainly astronomical for an institution our size with as many likely power users as we have. Considering the financial crisis we as an institution find ourselves in, and the attendant severe response by our administration, it is not hard to ask whether that money might be better spent, for example, reviving paused humanities graduate programs.
Anywhere in between these limiting scenarios, and the question becomes: If Anthropic dramatically increases prices in the future, will we be ready to cut ties, or will the institution be addicted to the product because our collective cognitive muscles have atrophied in the meantime?
The correct response to the disruption posed by LLMs is a realignment of our priorities and a reinvestment in the human experience of teaching and learning. Especially in the sciences, where we face an uncertain future for research funding and have a reputation for offering introductory courses that can be more like processed meat than prix fixe dining, we should face this moment not by surrendering to the bots, but by showing the students what world-class education really looks like.
Mark Levin is a professor of chemistry at the University of Chicago.
